JIT: optimize tiering policy, warmup, and compilation cost

[chrisbbreuer]

Parent: #459
Depends on: #461, #466, #467, #471, #480

Goal

Build a general tiering policy that minimizes time-to-performance, compile CPU, code size, invalidation churn, and warmup latency across short- and long-lived workloads.

Current boundary

The comparison harness deliberately performs ten warm calls to cross the current eight-entry optimizer threshold. That validates warmed performance but does not prove the threshold/policy is right for applications, cold work, polymorphism, shared realms, or large code.

Scope

• Profile entry/backedge heat, loop trip count, dispatch cost, runtime-operation mix, polymorphism, allocation, deopt/invalidation history, compile cost, and expected remaining lifetime.
• Select interpreter/VM/baseline/optimizer/OSR transitions with a bounded cost-benefit model rather than one global count.
• Add compile queue/background compilation only if artifact claims, Context teardown, no-GIL readers, W^X, memory pressure, and deterministic tests remain exact.
• Apply exponential backoff/cooldown to repeated rejection/deopt/invalidation based on reason, while allowing changed evidence to retry.
• Bound compiled bytes and prioritize/evict by measured value under Memory: bound heap, cache, and executable-code footprint #480.

No-workaround rules

• No benchmark/function/source recognition, magic threshold tuned only to Perf: outperform JavaScriptCore across the verified benchmark matrix #44, permanent JIT disable after one ordinary miss, hidden warmup, or compile work moved outside a declared cold timer.

Acceptance

• Bench: define the representative dependency-free performance matrix #460 reports time-to-tier, compile latency/CPU, code bytes, publications, rejects, deopts, invalidations, and steady attribution per workload.
• Policy improves a predeclared blend of cold, warm, short, long, monomorphic, polymorphic, and no-GIL work without row-level regression.
• Deterministic policy tests and forced-tier semantic differentials cover every transition and backoff state.
• Compilation/publication/invalidation/teardown races pass TSan and artifact lifetime gates.
• Full reference-host evidence includes both wall time and compile/code-memory cost.

Foundation

Tier records from #52/#431 and managed-baseline preference at df86d687.